European AI act
March 30
European AI act
- It’s primarily a product safety law, secondarily a fundamental rights protection.
- It’s seated on two existing laws.
Regulation is a type of law (issued by the EU) that offers a high level of harmonization (doesn’t leave member states with a lot of discretionary power). Doesn’t have to be translated into national laws. Immediately protects all EU citizens and applies across the EU.
Rita’s notes
• civil law - comprehensive, codified statutes— Stare decisis — judges are bound by previous judicial decisions— Derived from England (11th century) and used in the US, UK, Canada, Australia, and India.- dversarial; lawyers play a central role, while judges act as neutral referees. • common law-relies on judicial precedent and case-by-case decisions— Codified statutes — comprehensive, written legal codes covering all areas of law.- Inquisitorial; judges are actively involved in investigating the facts of the case. • AI systems definition- they want to overregulalte or underregulate. They dont want to accidentally penalize systems which are not even AI • definition- - build phase- machine based system, there is a hardware qualquer que seja, biological or mechanic (pombos, corpos, crispr),,software has gotten trained on input - what if the machine is part biological? - use phase- - varying levels of autonomy- has to have autonomy, not an if then - adaptiveness after deplyment, - can have it or not - explicit and implicit objectives, - output can be— predictions, content, recommendations, desicions - • why AI ACT— more present, more controversies,— its a product safety law- p roduct safety law— which products can be sold in european internal market, sets requirements for products to enter it, ( similar to general product safety regulation, and medicatl devices regulations) - improve internal market,, also uptake of trustworthy ai • Areas outside of scope of ai act— - military, defense, - can use if not in borders in case by case scenario, international organizations - models sole purpose use for researchand development - research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service (unless tested in real world)* - personal not professional use of models - free open source AI systems • dividing AI systems into risk levels - unnacceptable- prohibited- social scoring, facial recognition, dark pattern, manipulation - each point has extra things that must be fulfilled, — significant harm, unjustified, disproportional, unless terrorrism threats - subliminal manipulative deceptive techniques, exploiting vulnerable people (age disability), predictive policing, scraping facial recogntiion databases from online without permission, emotion recognition in home or school, biometric categorization (divide population), real time biometric identification of face or voice in public - critiques- if you open the way for those exceptions how do you control that they are not being used otherwise - high risk- conformity assessment- education, emplyment, justice, immigration, law - - - limited risk- transparency- chat bots, deep fakes, emotion recogniition - minimal risk- code of conduct- spam filters, video games • regulation*- type of law issued by the EU, that provides uniform discupline of all EU member states, does not need to be transformed into national laws, has immediate horizontal effect- gives automatic rights and obligations • lex specialis , lex generalis • okay, its not classified as high risk- rules about how it can be built, and how it can be used -(obligations for providers and users)
image.png
• must be clear about what risks there are, and the companies must assess what about of risk is acceptable, there is always risk • conformity assessment- standard is self accessment- in biometric cases there are also external assessments possible. if they lie ofc they have fines - companies can assess themselves if they are complying— but if there are biometric situations then they can do internal or not control
image.png
• conformity assessments for high risk AI- - risk management - data governemnt- good quality, no biases - technical documents - automatic logs of ai activities - transarency of information - human oversight - designed for accuracy, robustness, cybersecurity • deregulation- simplified laws, want the market to do their thing, pesticides, data, ai,… - example- if youhad a high risk you had to state it and register it in public doc, but now maybe dont need to state it anymore (this is no accepted fully yet) • very very dangerous AI is in course, but high risk AI rules are delayed- it was gonna start now but was postponed,, — rules only apply maybe to new AI and not ones already circulating